Freedom, Intelligence and Agent

1. The Role of Human “Freedom” and “Individuality” in the Evolution of Intelligence

From Escape from Freedom

From an evolutionary perspective, the history of humankind can be seen as a process of deepening individuation and freedom. Humanity’s departure from the pre-human stage marked the first step away from the compulsion of instinct, marking the beginning of our quest for freedom.

If instinct represents a special behavioral pattern determined by our neural structure, we can observe a clear trend across the animal kingdom: the lower an animal’s level of development, the more perfectly it adapts to nature, and the more rigidly its behavior is governed by instinct and reflex. Some insects, for example, build entire societies purely through innate programming.

Conversely, the higher an animal’s level of development, the greater its behavioral flexibility becomes, and the less complete its pre-wired structure is at birth. Humans sit at the extreme end of this spectrum as the most helpless species when born, yet the most capable of learning. Our adaptation to the world depends primarily on learning rather than instinct. As Fromm notes, “Instincts, even when not entirely lost, continually decline in the higher animals, especially in humans.”

When instinctive behavior loses its dominance, adaptation to nature is no longer automatic, and behavior begins to escape the limits of innate mechanisms, human existence truly begins. In this sense, to exist as a human is to be free.

However, this freedom is not only the positive freedom “to act”; it is also the negative freedom “from compulsion,” a liberation from the dictates of instinct.

A curious fact here is that consciousness itself is a form of finitude. If we were infinite, events could unfold without sequence, randomly and meaninglessly. Yet we live in a world where technology expands freedom faster than our minds can adapt to it, forcing us to confront our own finitude. Anxiety is the by-product of that collision.

When our individuality grows beyond anything found in the natural world, we begin to suffer a kind of spiritual pressure because freedom has a cost. It is an evolutionary gift that our biological systems have not yet learned to bear. To put it bluntly, anxiety is freedom and freedom is anxiety.

2. AGI, Intelligence, and Consciousness

From the perspective of intelligence, I believe humanity’s evolutionary path was correct because freedom, or individuality, served as a key catalyst.

If we hope for agents to evolve toward artificial general intelligence (AGI), we may need to abandon many of today’s fixed frameworks such as rigid workflows that make baby AIs appear immediately competent, much like antelopes that can run the moment they are born. Although such rigidity grants immediate functionality, it may also lock AI into an evolutionary branch that cannot truly flourish.

From Anthropic’s agent talk (source): “A truly intelligent agent has only one defining feature: autonomy. As models improve every few months, systems built around autonomous agents will naturally advance. In contrast, heavily structured workflows restrict the model’s potential and prevent it from fully leveraging the next generation of intelligence. Many frameworks have become bloated and overly subjective. Some say, all you really need is a while loop. To build real intelligence, we must liberate the model and let autonomy become the core.

This idea resonates with the recent surge of reinforcement learning (RL) in 2025. Its success owes much to strong priors established by pre-training and supervised fine-tuning (SFT). Just as humans, though freer than other animals, still inherit genetic priors that help us learn quickly, RL benefits from a well-trained base. Earlier failures in RL stemmed from bases too weak to learn effectively in complex environments.

Yet it is worth asking whether deeper imitation learning and supervised learning are truly beneficial. Has the base model reached sufficient maturity to support genuine autonomy, or will overtraining push AI further toward the “antelope path,” efficient yet evolutionarily stagnant?

From the perspective of consciousness, I do not believe LLMs possess it because they lack finitude. Each model, of course, is finite in computation and parameters, but silicon-based systems have a form of relative infinity. Their weights can be stored, duplicated, updated, and redistributed indefinitely. As long as AI development continues, these systems effectively persist. Although one could raise many counter-examples, compared with carbon-based humans, silicon-based AIs inhabit a world of relative infinity.

However, when we examine the relationship between consciousness and intelligence, we find they are not tightly correlated. Intelligence can exist without consciousness, but without consciousness, intelligence struggles to produce self-motivation. It can reason, but it cannot care, and without caring, it is difficult to self-evolve.

3. On Self-Evolving Agents

How, then, can a non-conscious LLM become a self-evolving agent? This question has fascinated me lately.

Two aspects feel particularly important:

1. How to move beyond explicit rewards and engage with the real world through implicit signals;
2. How to map experience back into the model itself.

1. Beyond explicit reward In the real world, 99% of actions do not yield a clear reward. Future LLMs will eventually have to graduate from their school life, leaving behind benchmarks, exams, and explicit scores, and enter adulthood. In this stage, “baby intelligence” will discover that most of reality is ambiguous, long-term, and chaotic. There is no easy feedback, no immediate validation, and no clean metric of success.

2. Mapping experience back into the model We know that an LLM’s weights encode what it has learned from prior training. Once an agent interacts with the real world and gathers new insights, the question becomes how this new experience can be integrated gracefully and meaningfully into its own parameters so that it can evolve. Many approaches attempt this by extending memory into external databases or by directly adjusting the weights (ΔW) after new experiences, yet none have found the optimal solution.